Generated with PaperBanana (Gemini 3 Pro). Shows query expansion
fanning HyDE+Vec into vector searches, Lex into BM25, merged via
reciprocal rank fusion and LLM reranking.
List query first in --help as the recommended search method. Add
vector-search and deep-search as undocumented CLI aliases matching
MCP tool names.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
searchResultsToMarkdown and searchResultsToXml in formatter.ts were
silently dropping the context field. Added formatter.test.ts covering
context visibility across all output formats.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* feat: MCP HTTP transport with daemon lifecycle
Add streaming HTTP transport as an alternative to stdio for the MCP
server. A long-lived HTTP server avoids reloading 3 GGUF models (~2GB)
on every client connection, reducing warm query latency from ~16s (CLI)
to ~10s.
New CLI surface:
qmd mcp --http [--port N] # foreground, default port 3000
qmd mcp --http --daemon # background, PID in ~/.cache/qmd/mcp.pid
qmd mcp stop # stop daemon via PID file
qmd status # now shows MCP daemon liveness
Server implementation (mcp.ts):
- Extract createMcpServer(store) shared by stdio and HTTP transports
- HTTP transport uses WebStandardStreamableHTTPServerTransport with
JSON responses (stateless, no SSE)
- /health endpoint with uptime, /mcp for MCP protocol, 404 otherwise
- Request logging to stderr with timestamps, tool names, query args
Daemon lifecycle (qmd.ts):
- PID file + log file management with stale PID detection
- Absolute paths in Bun.spawn (process.execPath + import.meta.path)
so daemon works regardless of cwd
- mkdirSync for cache dir on fresh installs
- Removes top-level SIGTERM/SIGINT handlers before starting HTTP
server so async cleanup in mcp.ts actually runs
Move hybridQuery() and vectorSearchQuery() into store.ts as standalone
functions that take a Store as first argument. Both CLI and MCP now
call the identical pipeline, eliminating the class of bugs where one
copy drifts from the other.
Shared pipeline (store.ts):
- hybridQuery(): BM25 probe → expand → FTS+vec search → RRF →
chunk → rerank (chunks only) → position-aware blending → dedup
- vectorSearchQuery(): expand → vec search → dedup → sort
- SearchHooks interface for optional progress callbacks
- Constants: STRONG_SIGNAL_MIN_SCORE, STRONG_SIGNAL_MIN_GAP,
RERANK_CANDIDATE_LIMIT (40), addLineNumbers()
Bugs fixed by unification:
- MCP now gets strong-signal short-circuit (was CLI-only)
- Reranker candidate limit unified at 40 (MCP had 30)
- File dedup added to hybrid query (MCP was missing it)
- Collection filter pushed into searchVec DB query
- Filter-then-slice ordering fixed (MCP was slice-then-filter)
* feat: type-routed query expansion — lex→FTS, vec/hyde→vector
expandQuery() now returns typed ExpandedQuery[] instead of string[],
preserving the lex/vec/hyde type info from the LLM's GBNF-structured
output. hybridQuery() and vectorSearchQuery() route searches by type:
lex queries go to FTS only, vec/hyde go to vector only.
Previously, every expanded query ran through BOTH backends — keyword
variants wasted embedding forward passes, semantic paraphrases wasted
BM25 lookups. Type routing eliminates ~4 calls/query with zero quality
loss (cross-backend noise actually hurt RRF fusion).
Cache format changed from newline-separated text to JSON (preserves
types). Old cache entries gracefully re-expand on first access.
CLI expansion tree now shows query types:
├─ original query
├─ lex: keyword variant
├─ vec: semantic meaning
└─ hyde: hypothetical document...
Benchmark (5 queries, 1756-doc index, warm LLM, Apple Silicon):
Metric Old (untyped) New (typed) Delta
Avg backend calls 10.0 6.0 -40%
Total wall time 1278ms 549ms -57%
Avg saved/query — — 146ms
"authentication setup" 12 → 7 calls 511 → 112ms
"database migration strategy" 10 → 6 calls 182 → 106ms
"how to handle errors in API" 10 → 6 calls 216 → 121ms
"meeting notes from last week" 10 → 6 calls 228 → 110ms
"performance optimization" 8 → 5 calls 141 → 100ms
Savings come from skipped embed() calls (~30-80ms each). FTS is
synchronous SQLite (~0ms), so lex→FTS routing is free while
vec/hyde→vector-only avoids wasted embedding passes.
* fix: MCP query snippets now use reranker's best chunk, not full body
extractSnippet() was scanning the entire document body for keyword
matches to build the snippet. But hybridQuery() already identified
the most relevant chunk via cross-attention reranking — rescanning
the full body is redundant and can land on a less relevant section
if the query terms appear elsewhere in the document.
CLI was already using bestChunk (set during the refactor). MCP was
still using body — a pre-existing inconsistency, not a regression.
* feat: dynamic MCP instructions + tool annotations
The MCP server now generates instructions at startup from actual index
state and injects them into the initialize response. LLMs see collection
names, document counts, content descriptions, and search strategy
guidance in their system prompt — zero tool calls needed for orientation.
Previously, the only guidance was generic static tool descriptions and
a user-invocable "query" prompt that no LLM would discover on its own.
An LLM connecting to QMD had no idea what collections existed, what they
contained, or how to scope searches effectively.
* change default port to 8181
* fix: BM25 score normalization was inverted
The normalization formula `1 / (1 + |bm25|)` is a decreasing function of
match strength. FTS5 BM25 scores are negative where more negative = better
match (e.g., -10 is strong, -0.5 is weak). The formula mapped:
strong match (raw -10) → 1/(1+10) = 9% ← should be highest
weak match (raw -0.5) → 1/(1+0.5) = 67% ← should be lowest
Three downstream effects:
1. `--min-score 0.5` (or MCP minScore: 0.5) filtered OUT strong matches
and kept only weak ones. The MCP instructions recommend this threshold.
2. CLI `formatScore()` color bands never showed green for BM25 results
(best matches scored ~9%, green threshold is 70%).
3. The strong signal optimization in hybridQuery (skip ~2s LLM expansion
when BM25 already has a clear winner) was dead code — strong matches
scored ~0.09, never reaching the 0.85 threshold.
Fix: `|x| / (1 + |x|)` — same (0,1) range, monotonic, no per-query
normalization needed, but now correctly maps strong → high, weak → low.
The normalization was born broken (Math.max(0, x) clamped all
negative BM25 to 0 → every score = 1.0), then PR #76 changed to
Math.abs which made scores vary but inverted the direction. Neither
state was ever correct.
* fix: rerank cache key ignores chunk content
The rerank cache key was (query, file, model) but the actual text sent
to the reranker is a keyword-selected chunk that varies by query terms.
Two different queries hitting the same file can select different chunks,
but the second query gets a stale cached score from the first chunk.
Example:
Query "auth flow" → selects chunk about authentication → score 0.92
Query "auth tokens" → same file, selects chunk about tokens
→ cache HIT on (query, file, model) → returns 0.92 from wrong chunk
Fix: include full chunk text in cache key. getCacheKey() already
SHA-256 hashes its inputs, so this adds no key bloat — just
disambiguation. Old cache entries become natural misses (different key
shape) and re-warm on next query.
* rename MCP tools for clarity, rewrite descriptions for LLM tool selection
Rename MCP tools: vsearch → vector_search, query → deep_search.
LLMs see these names — self-documenting names reduce reliance on
descriptions for tool selection. CLI commands stay unchanged
(qmd vsearch, qmd query) — different namespace, users type those.
Rewrite all search tool descriptions to be action-oriented:
- search: "Search by keyword. Finds documents containing exact
words and phrases in the query."
- vector_search: "Search by meaning. Finds relevant documents even
when they use different words than the query — handles synonyms,
paraphrases, and related concepts."
- deep_search: "Deep search. Auto-expands the query into variations,
searches each by keyword and meaning, and reranks for top hits
across all results."
Rewrite instructions ladder — each tool says what it does, no
"start here" / "escalate as needed" strategy language.
Delete the "query" prompt (registerPrompt) — it restated what
descriptions + instructions already cover. No LLM proactively
calls prompts/get to learn how to use tools.
* supress HTTP server logs during tests
- Add marketplace.json for Claude Code plugin installation
- Simplify skill status check to inline `qmd status` (portable across agents)
- Update SKILL.md MCP section, reference mcp-setup.md for manual config
- Clean up mcp-setup.md (remove redundant prerequisites)
- Rename MCP-SETUP.md to mcp-setup.md
Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
BM25 scores in SQLite FTS5 are negative (lower = better match).
The previous code used Math.max(0, score) which clamped all negative
scores to 0, resulting in all results showing 100% (score = 1.0).
Fix: Use Math.abs(score) to properly convert negative BM25 scores
to positive values for the normalization formula.
Before: All results show Score: 100%
After: Scores vary based on actual BM25 relevance (e.g., 16%, 5%, 6%)
Fixes#74
Replace Bun.file() async calls with Node.js fs sync methods to work
around a Bun bug that corrupts UTF-8 file paths containing non-ASCII
characters.
Bug: Bun.file(filepath).stat() and Bun.file(filepath).text() internally
mangle UTF-8 encoding, causing ENOENT errors with mojibake paths when
accessing files in iCloud Drive and other locations.
Changes:
- src/qmd.ts: Use readFileSync instead of Bun.file().text()
- src/qmd.ts: Use statSync instead of Bun.file().stat() for file metadata
- src/store.ts: Use statSync for SQLite custom path detection
* Fix: Add missing --index option to argument parser
The --index flag was documented and used in code but not defined
in parseArgs options, causing it to be ignored. Now properly handles
custom index names like: qmd --index test status
* Feature: Use index name for config files too
Now --index <name> loads ~/.config/qmd/<name>.yml instead of index.yml.
This allows completely separate indexes with their own collections.
Example:
qmd --index hackage status
→ Uses ~/.config/qmd/hackage.yml + ~/.cache/qmd/hackage.sqlite
Moved hackage collection to hackage.yml for separation.
- Changed temperature from 0/0.1 to 0.7 (Qwen3 non-thinking mode default)
- Added topK=20, topP=0.8 per Qwen3 docs
- Added repeatPenalty with presencePenalty=0.5 for query expansion
- Fixes infinite loop on acronyms like DHH, BFCM
Qwen3 docs explicitly warn: 'DO NOT use greedy decoding, as it can
lead to performance degradation and endless repetitions'
- List all HuggingFace repos in CLAUDE.md (model, gguf, sft, grpo, train)
- Update jobs scripts to use tobil/qmd-query-expansion-train (no -v2)
- Clarify rules: no versioned repos, update in place
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Remove versioned files (sft_v4.yaml, prepare_v4_dataset.py, train_v2/)
- Update configs to use local data/train/ directory
- Add glob pattern support to prepare_data.py and train.py
- Update .gitignore to properly ignore outputs/ and data/train*/
- Document data preparation step in CLAUDE.md
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add finetune/CLAUDE.md documenting the training pipeline
- Update configs to output to local outputs/ directory (gitignored)
- Document that all data/*.jsonl files are training data
- Document local CUDA training vs HuggingFace Jobs cloud training
- Enforce eval requirement before any model upload
- Single model repo (no -v1, -v2, -v4 versioning)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Brings in:
- /only: variants for single-type expansions
- LLM session management for lifecycle safety
- skills.sh integration for AI agent discovery
- Various bug fixes for vector search and embeddings
Merge conflicts resolved by keeping hyde-first format ordering
from finetune branch while accepting expanded templates and
new features from main.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Move the hyde (hypothetical document) line to the beginning of the
output format, before lex and vec lines. This better reflects the
logical flow where the hypothetical document is generated first and
then informs the keyword/semantic expansions.
Also adds auto-download of eval_common.py in training scripts for
standalone HuggingFace Jobs execution.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- generate_only_variants.py: Creates training data where queries end with
'only: lex', 'only: vec', or 'only: hyde' and output contains ONLY that type
- reward.py: Updated scorer to handle 'only:' mode separately
- Penalizes presence of unwanted types
- Type-specific quality checks
- Filters templated low-quality hyde outputs
- 4,444 high-quality 'only:' variants from v2 + handcrafted data
Adds a session layer that prevents LLM contexts from being disposed
mid-operation during long-running tasks like batch embedding or
multi-step search workflows (expand → embed → rerank).
Key changes:
- Add LLMSessionManager with reference counting for active sessions
- Add LLMSession class for scoped access with automatic acquire/release
- Add withLLMSession() API for multi-step workflows
- Update idle timer to check canUnloadLLM() before disposing
- Wrap querySearch, vectorSearch, and embed command in sessions
- Add optional session parameter to searchVec and getEmbedding
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* Add query expansion model finetuning infrastructure
- Training scripts for Qwen3-0.6B and 1.7B models
- Dataset generation from s-emanuilov/query-expansion
- Evaluation scripts comparing finetuned vs baseline models
- GRPO RL training script (optional improvement)
- Export script for GGUF conversion
Results:
- 0.6B finetuned: 95% format compliance (lex/vec/hyde)
- Baseline: 0% format compliance
- Dataset: 5,157 examples on HuggingFace Hub
Models available at:
- tobil/qmd-query-expansion-0.6B (recommended)
- tobil/qmd-query-expansion-train (dataset)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* Fix GRPO training script for TRL API compatibility
- Use max_completion_length instead of max_new_tokens
- Use processing_class instead of tokenizer
- Use args instead of config for GRPOTrainer
- Add __name__ attribute to reward function class
- Accept **kwargs in reward function for extra TRL args
- Add new LoRA adapter after merging SFT weights
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* Update README with final evaluation results
- 0.6B SFT: 95% format compliance (best)
- 0.6B GRPO: 0% (catastrophic forgetting from RL)
- 1.7B v2: training completed, evaluation pending
- Added GRPO evaluation results
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* Add comprehensive scoring system for query expansion
New scoring criteria (0-100 points):
- Format (30): Must have lex: and vec: prefixes
- Diversity (30): Multiple types, no echoing query, diverse expansions
- Hyde (20): Optional, concise, no newlines, no word repetition
- Quality (20): Lex=keywords, vec=natural language
See SCORING.md for full documentation.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* Add HuggingFace login and comprehensive scoring to GRPO v2 training
- Add explicit HF_TOKEN login before training
- Use SCORING.md criteria as RL reward function
- Conservative training: LR 1e-6, LoRA rank 4
- Reward scores: good=0.94, bad=0.38
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* Refactor finetune folder: train/rl scripts with YAML configs
Major changes:
- train.py: Generic SFT training script using YAML config
- rl.py: Generic GRPO training script using YAML config
- configs/: YAML configs per training run (sft_v4.yaml, grpo_v4.yaml)
- dataset/: Data preparation scripts moved here
- tui.py: Interactive model testing interface
Training results:
- SFT v4: 98.8% avg score (all Excellent)
- GRPO v4: 0% (failed - model drifted to verbose explanations)
Removed per-model scripts (train_0.6B.py, train_1.7B.py, etc)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* Add named entity extraction to GRPO reward function
Key changes:
- Extract named entities (acronyms, proper nouns, technical terms)
- Heavy penalty (-30) when lex queries miss named entities
- Penalty (-15) for generic filler phrases like "find information about"
- Compound entity detection (TDS motorsports -> both words)
- Update GRPO config with KL regularization (beta=0.04)
- Lower learning rate (5e-7) and add max_steps (200)
Test results:
- "who is TDS motorsports" good: 1.00, bad: 0.30 (was 0.75)
- "how to use React hooks" good: 0.87, bad: 0.45 (was 0.75)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* Add chat template leakage detection to reward function
Zero reward for outputs containing:
- <|im_start|>, <|im_end|> tokens
- <think>, </think> tags (Qwen3 thinking mode)
- Role markers like \nassistant\n, \nuser\n
- <|endoftext|> token
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* Strict format validation: every line must be lex:/vec:/hyde:
Any line that doesn't start with a valid prefix now returns 0.0
instead of just counting as a penalty. This prevents any prose,
explanations, bullet points, or other invalid content.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* Clean up evaluation files
- Remove old versioned evaluation files (0.6B, 1.7B, baseline)
- Rename evaluation_v4.json -> evaluation_sft.json
- Rename evaluation_v4_grpo.json -> evaluation_grpo_failed.json
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* Refactor evals into separate run and score scripts
New structure:
- evals/run.py: Generate model outputs to JSONL
- evals/score.py: Score outputs with detailed breakdown
- evals/queries.txt: Test queries (26 total)
Features:
- Supports both HF Hub and local model paths
- Named entity preservation scoring
- Chat template leakage detection
- Strict format validation (every line must be lex:/vec:/hyde:)
- Generic phrase detection
Usage:
uv run evals/run.py --model tobil/qmd-query-expansion-0.6B-v4
uv run evals/score.py evals/results_*.jsonl
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* Fix GRPO model loading to use SFT base first
The GRPO adapter was trained on merged SFT weights, so loading it
directly on the base model results in 0% score. Added --sft-model
parameter to evals/run.py to load SFT first, then apply GRPO adapter.
With correct loading: GRPO scores 89.7% (all 26 queries Excellent).
Updated README with correct GRPO score and loading instructions.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* Fix TUI to load GRPO models with SFT base first
GRPO adapters were trained on merged SFT weights, so they need SFT
loaded and merged first before applying the GRPO adapter.
Updated MODELS config to include sft_base path for GRPO models,
and load_model() now handles the SFT -> merge -> GRPO flow.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* Update README for unified model repository structure
All models (0.6B, 1.7B, 4B) with SFT and GRPO variants now go into
a single HuggingFace repo (tobil/qmd-query-expansion) with subfolders
for each size and training method.
Updated loading examples to show subfolder-based model loading.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* Update README with separate model repos
Changed from subfolder approach to separate repos per model since
trainer.push_to_hub() doesn't support subfolder argument.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* Add 1.7B and 4B GRPO training and GGUF conversion scripts
Training scripts for GRPO fine-tuning:
- train_1.7B_grpo.py: GRPO training for Qwen3-1.7B
- train_4B_grpo.py: GRPO training for Qwen3-4B
GGUF conversion scripts:
- convert_1.7B_gguf.py: Merge SFT+GRPO adapters and convert to GGUF
- convert_4B_gguf.py: Merge SFT+GRPO adapters and convert to GGUF
All scripts use PEP 723 inline dependencies for HuggingFace Jobs.
Models published:
- tobil/qmd-query-expansion-1.7B-sft
- tobil/qmd-query-expansion-1.7B-grpo
- tobil/qmd-query-expansion-1.7B-gguf
- tobil/qmd-query-expansion-4B-sft
- tobil/qmd-query-expansion-4B-grpo
- tobil/qmd-query-expansion-4B-gguf
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* Remove beads issue tracking
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* Remove beads reference from CLAUDE.md
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* Fix GRPO reward function to handle think blocks and end tokens
- Strip <|im_end|> token from completions (model output includes it)
- Change think_penalty to skipped_think bonus (+20 for not using think)
- Adjust max_possible to account for bonus (120/140)
- Fix typo in chat template artifact check
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* Make TUI model list dynamic from HuggingFace Hub
- Fetch available qmd-query-expansion models from tobil/ on Hub
- Auto-detect model size (0.6B, 1.7B, 4B) and use correct base model
- Group models by type (SFT vs GRPO) in menu
- Skip GGUF repos in model listing
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* Fix GRPO training: apply chat template to prompts
The SFT model was trained with chat template format but GRPO was
passing raw prompts. Now prompts are formatted with tokenizer.apply_chat_template()
so the model sees the same format it learned during SFT.
Also update extract_query_from_prompt to strip chat template artifacts.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* Finetune 2.0: consolidate and simplify the entire training pipeline
Consolidate ~2,800 lines of duplicated code across 12 files into 5 clean,
well-documented files targeting Qwen3-1.7B end-to-end.
Key changes:
- Extract reward function into single source of truth (reward.py)
Previously duplicated 3x with divergent bugs across rl.py,
train_1.7B_grpo.py, and train_4B_grpo.py
- Unify training into one script with sft/grpo subcommands (train.py)
Replaces train.py + rl.py + train_1.7B_grpo.py + train_4B_grpo.py
- Merge eval generate+score into single eval.py
Replaces evals/run.py + evals/score.py
- Parameterize GGUF conversion by --size (convert_gguf.py)
Replaces convert_1.7B_gguf.py + convert_4B_gguf.py
- Fix critical bug: rl.py silently ignored beta/temperature from config,
causing the exact catastrophic drift its own comments warned about
- Fix prompt consistency: all files use /no_think chat template format
- Retarget configs from 0.6B to 1.7B
- Comprehensive README documenting the full pipeline
Removed: rl.py, train_1.7B_grpo.py, train_4B_grpo.py, convert_1.7B_gguf.py,
convert_4B_gguf.py, tui.py, evals/run.py, evals/score.py
Net: -3,429 lines, +382 lines
Co-Authored-By: Claude (claude-fudge-eap-cc) <noreply@anthropic.com>
* Add HF Jobs scripts, temporal query examples, and training results
- jobs/sft.py and jobs/grpo.py: self-contained scripts for
`hf jobs uv run` (no local GPU needed)
- 12 temporal/recency query examples in training data (e.g. "recent
news about Shopify" -> lex with years 2025/2026)
- 4 temporal test queries in evals/queries.txt
- README updated with HF Jobs workflow, training results, and
updated file structure
- Remove .beads tracking
SFT and GRPO successfully trained on A10G via HF Jobs:
SFT: eval loss 0.321, token accuracy 92.4%
GRPO: mean reward 0.757, 200 steps, KL 0.00048
Co-Authored-By: Claude (claude-fudge-eap-cc) <noreply@anthropic.com>
* Deploy fine-tuned GRPO model as default for query expansion
Switch from generic Qwen3-1.7B-Q8_0 (~2.2GB) to fine-tuned
qmd-query-expansion-1.7B-q4_k_m (~1.1GB). The fine-tuned Q4
scores 91.7% avg with 30/30 Excellent, outperforming the base Q8.
- Update default generate model in src/llm.ts
- Update README model table, architecture diagram, config block
- Add v2 training data, eval scripts, and quantize job
- Remove superseded v1 training data (5,742 → 1,000 examples)
- Update finetune README with v2 results and file structure
Co-Authored-By: Claude (claude-fudge-eap-cc) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
Switch from generic Qwen3-1.7B-Q8_0 (~2.2GB) to fine-tuned
qmd-query-expansion-1.7B-q4_k_m (~1.1GB). The fine-tuned Q4
scores 91.7% avg with 30/30 Excellent, outperforming the base Q8.
- Update default generate model in src/llm.ts
- Update README model table, architecture diagram, config block
- Add v2 training data, eval scripts, and quantize job
- Remove superseded v1 training data (5,742 → 1,000 examples)
- Update finetune README with v2 results and file structure
Co-Authored-By: Claude (claude-fudge-eap-cc) <noreply@anthropic.com>
- jobs/sft.py and jobs/grpo.py: self-contained scripts for
`hf jobs uv run` (no local GPU needed)
- 12 temporal/recency query examples in training data (e.g. "recent
news about Shopify" -> lex with years 2025/2026)
- 4 temporal test queries in evals/queries.txt
- README updated with HF Jobs workflow, training results, and
updated file structure
- Remove .beads tracking
SFT and GRPO successfully trained on A10G via HF Jobs:
SFT: eval loss 0.321, token accuracy 92.4%
GRPO: mean reward 0.757, 200 steps, KL 0.00048
Co-Authored-By: Claude (claude-fudge-eap-cc) <noreply@anthropic.com>
Consolidate ~2,800 lines of duplicated code across 12 files into 5 clean,
well-documented files targeting Qwen3-1.7B end-to-end.
Key changes:
- Extract reward function into single source of truth (reward.py)
Previously duplicated 3x with divergent bugs across rl.py,
train_1.7B_grpo.py, and train_4B_grpo.py
- Unify training into one script with sft/grpo subcommands (train.py)
Replaces train.py + rl.py + train_1.7B_grpo.py + train_4B_grpo.py
- Merge eval generate+score into single eval.py
Replaces evals/run.py + evals/score.py
- Parameterize GGUF conversion by --size (convert_gguf.py)
Replaces convert_1.7B_gguf.py + convert_4B_gguf.py
- Fix critical bug: rl.py silently ignored beta/temperature from config,
causing the exact catastrophic drift its own comments warned about
- Fix prompt consistency: all files use /no_think chat template format
- Retarget configs from 0.6B to 1.7B
- Comprehensive README documenting the full pipeline
Removed: rl.py, train_1.7B_grpo.py, train_4B_grpo.py, convert_1.7B_gguf.py,
convert_4B_gguf.py, tui.py, evals/run.py, evals/score.py
Net: -3,429 lines, +382 lines
Co-Authored-By: Claude (claude-fudge-eap-cc) <noreply@anthropic.com>
* fix: add promise guard to ensureEmbedContext to prevent race condition
Root cause: ensureEmbedContext() was not thread-safe. When multiple parallel
embedding requests called ensureEmbedContext() simultaneously, all would see
embedContext === null and start creating new contexts. This race condition
caused 'Context is disposed' errors as contexts were overwritten/orphaned.
The fix adds a promise guard (embedContextCreatePromise) to ensure only one
context creation runs at a time - identical to the pattern already used in
ensureGenerateModel().
Changes:
- Add embedContextCreatePromise field to track in-progress context creation
- Modify ensureEmbedContext() to wait for existing creation if in progress
- Update test comment and timeout for CPU-only systems
Testing:
- Fresh model download + qmd embed: 28/28 chunks succeeded (was 14/27)
- All embedBatch tests pass
- No warmup hack needed - full parallel performance from the start
Environment tested:
- Ubuntu 24.04 LTS (x64), Bun 1.3.6, node-llama-cpp 3.14.5, no GPU
* test: improve race condition test to verify single context creation
The previous test only verified embeddings succeeded but didn't prove the fix
actually prevents multiple context creation. This improved test:
- Instruments createEmbeddingContext to count invocations
- Runs 5 concurrent embedBatch calls on a fresh LlamaCpp instance
- Asserts exactly 1 context is created (fails with 5 without the fix)
Verified locally:
- With fix: 1 context created (PASS)
- Without fix: 5 contexts created (FAIL)
* chore: clear embedContextCreatePromise in dispose() for consistency
Refactor extractTitle to use extension-based extractors:
- .md: preserves original markdown logic (Notes skip behavior)
- .org: extracts from #+TITLE: property or first * heading
Extensions are lowercased for case-insensitive matching.
Easy to add more file types in the future.
Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
Override sqlite in devShell to enable extension loading for sqlite-vec
support when running tests. Only sets BREW_PREFIX if not already defined
to avoid overriding user's existing setup.
Package build remains unchanged.
Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
The SFT model was trained with chat template format but GRPO was
passing raw prompts. Now prompts are formatted with tokenizer.apply_chat_template()
so the model sees the same format it learned during SFT.
Also update extract_query_from_prompt to strip chat template artifacts.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Fetch available qmd-query-expansion models from tobil/ on Hub
- Auto-detect model size (0.6B, 1.7B, 4B) and use correct base model
- Group models by type (SFT vs GRPO) in menu
- Skip GGUF repos in model listing
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Strip <|im_end|> token from completions (model output includes it)
- Change think_penalty to skipped_think bonus (+20 for not using think)
- Adjust max_possible to account for bonus (120/140)
- Fix typo in chat template artifact check
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Training scripts for GRPO fine-tuning:
- train_1.7B_grpo.py: GRPO training for Qwen3-1.7B
- train_4B_grpo.py: GRPO training for Qwen3-4B
GGUF conversion scripts:
- convert_1.7B_gguf.py: Merge SFT+GRPO adapters and convert to GGUF
- convert_4B_gguf.py: Merge SFT+GRPO adapters and convert to GGUF
All scripts use PEP 723 inline dependencies for HuggingFace Jobs.
Models published:
- tobil/qmd-query-expansion-1.7B-sft
- tobil/qmd-query-expansion-1.7B-grpo
- tobil/qmd-query-expansion-1.7B-gguf
- tobil/qmd-query-expansion-4B-sft
- tobil/qmd-query-expansion-4B-grpo
- tobil/qmd-query-expansion-4B-gguf
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Changed from subfolder approach to separate repos per model since
trainer.push_to_hub() doesn't support subfolder argument.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
All models (0.6B, 1.7B, 4B) with SFT and GRPO variants now go into
a single HuggingFace repo (tobil/qmd-query-expansion) with subfolders
for each size and training method.
Updated loading examples to show subfolder-based model loading.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
GRPO adapters were trained on merged SFT weights, so they need SFT
loaded and merged first before applying the GRPO adapter.
Updated MODELS config to include sft_base path for GRPO models,
and load_model() now handles the SFT -> merge -> GRPO flow.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The GRPO adapter was trained on merged SFT weights, so loading it
directly on the base model results in 0% score. Added --sft-model
parameter to evals/run.py to load SFT first, then apply GRPO adapter.
With correct loading: GRPO scores 89.7% (all 26 queries Excellent).
Updated README with correct GRPO score and loading instructions.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
