Resolve conflicts: combine AST chunking args (filepath, chunkStrategy)
with abort signal parameter from #458.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add missing subprocess import (NameError on any quantize path)
- Replace broken optimum-cli quantize calls with direct onnxruntime:
Q4 uses MatMulNBitsQuantizer, Q8 uses quantize_dynamic
- Add onnxconverter-common to deps for FP16 (was silently swallowed)
- Make FP16 fail loudly on missing dep instead of silently uploading FP32
- README and transformers_js_config now reflect actual quantize_type
instead of always hardcoding Q4
- Remove dead _convert_fp16_external function
- Use no_post_process=True for ONNX export to avoid protobuf serialize error
- Add --validate and --validate-only flags for inference verification
- Fix position_ids in validation feed (required by Qwen3 ONNX export)
- Use optimum-cli for quantization to handle external data format
- Fix optimum dependency to optimum[onnxruntime]
Tested: export + validation passes on CPU, KV cache present (56 tensors).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add convert_onnx.py that mirrors convert_gguf.py's structure:
- Loads base Qwen3 model, merges SFT + GRPO adapters
- Exports to ONNX via Optimum (text-generation-with-past task)
- Supports Q4 (MatMulNBits), Q8, FP16, and FP32 output
- Uploads to separate HF repo (e.g. tobil/qmd-query-expansion-1.7B-ONNX)
- Writes Transformers.js compatibility config
- Includes model card with usage example
Usage:
uv run convert_onnx.py --size 1.7B
uv run convert_onnx.py --size 1.7B --quantize q4 --no-upload
Also adds `just convert-onnx` and `just convert-gguf` tasks.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Compound entity chaining now stops one level deep. Previously "TDS
motorsports team history" would inflate the expected entity set with
"team" and "history", causing false-positive entity-preservation
penalties during GRPO. Now only {tds, motorsports} are detected.
- Add INTERIOR_FILLER_WORDS penalty (-3/line): lex lines containing
"overview" or "basics" absent from the original query are penalised.
Targets template-generator noise, e.g. "ancient overview rome timeline".
- Raise is_diverse threshold 2→3: requires 3 unique words between lex
lines before they count as diverse. Reduces reward for near-duplicate
pairs like "auth setup" / "auth configuration".
- Broaden quoted-phrase bonus: was gated on named entities existing;
now any multi-word query earns +3 for using quotes in lex lines.
Better incentivises BM25-aware syntax like "memory leak" python.
Fixes scoring noise identified while working on issue #247.
Replace ad-hoc JSON parsing with a strict Pydantic model
(TrainingExample with typed OutputPair). All data loading goes
through load_examples() which fails loudly on invalid data.
- Convert v3_structured.jsonl from "searches" to "output" format
- Rewrite all consumer scripts (prepare, validate, score, analyze)
to load through the Pydantic schema
- Prepared train/val files are ephemeral build artifacts
- Restore LFM2 and GEPA experiments under experiments/
- Add pydantic>=2.0 to dependencies
Training data:
- Expand lex phrases/negation examples from 12 to 74 with intent field
- Add 50 personal entity examples (meetings, emails, projects with names)
Reward function:
- Detect entities at position 0 (fixes "Bob asked about deploy")
- Per-entity coverage penalty: -20 per entity absent from all lex+vec
- Phrase quoting bonus: +3 when lex uses quotes for multi-word terms
- Expanded stopwords to reduce false positive entity detection
Eval queries: add 21 test queries for personal entities, quoted phrases,
and negation/disambiguation scenarios.
Remove one-off data generator/fix scripts, superseded data files (v2, v3
replaced by v3_structured), LFM2 experiment, GEPA directory, duplicate
job scripts, and historical docs. Clean up Justfile.
These are restored under experiments/ in a later commit.
The 'query document' is now a first-class concept in QMD: a structured
document with typed sub-queries that combine for best recall.
## Query types
- lex: BM25 keyword search with phrase and negation syntax
- vec: Semantic vector search (natural language questions)
- hyde: Hypothetical document (write the expected answer)
- expand: Auto-expand via local LLM (max 1, default for plain queries)
## Lex syntax
Full BM25 operator support:
"exact phrase" verbatim match, no prefix
-term exclude documents containing term
-"exact phrase" exclude documents containing phrase
Examples:
"C++ performance" optimization -sports -athlete
"connection pool" timeout -redis
"machine learning" -sports -athlete
## MCP tool description rewritten
The 'query' tool description now fully teaches AI agents the query
document format, lex syntax, and strategy for combining types.
Includes worked examples including intent-aware lex (C++ performance,
not sports) which is critical for disambiguation in dense corpora.
## Unit tests
11 new lex parser tests covering:
- plain terms, quoted phrases, negation, combined
- intent-aware disambiguation (performance -sports -athlete)
- only-negation returns null (FTS5 constraint)
- empty/whitespace handling
## Training data
12 new intent-aware examples for next model training round:
- Real technical topics with lex phrase+negation combinations
- Covers: C++ perf, Python memory, DB connections, rate limiting,
SQL optimization, ML overfitting, Docker, JWT, async/await,
git conflicts, Kubernetes, React state
- Each shows how context/intent shapes lex query construction
(e.g. performance with C++ context → -sports -athlete exclusions)
Dataset improvements:
- Reorder output to put hyde first for better retrieval priming
- Convert absolute paths to relative paths in scripts
- Add convert_to_structured.py for structured data format
- Add qmd_expansion_v3_structured.jsonl with type/query objects
- Update schema.py with reorder_hyde_first() helper
- Verify data now validates hyde-first ordering
Training data regenerated with new ordering (100% validation success).
Add training configuration and documentation for using LiquidAI's LFM2-1.2B
as an alternative base model for query expansion fine-tuning.
LFM2 benefits:
- 2x faster decode/prefill vs standard transformers
- Optimized for edge/on-device inference
- Good at agentic tasks, RAG, and data extraction
Changes:
- Add configs/sft_lfm2.yaml with LFM2-specific LoRA target modules
- Add jobs/sft_lfm2.py for HuggingFace Jobs training
- Update llm.ts with LFM2 GGUF model URIs
- Add documentation for LFM2 training workflow
LFM2 uses a hybrid architecture (convolutions + attention) requiring
different LoRA targets: q_proj, k_proj, v_proj, out_proj, in_proj, w1, w2, w3
Dataset improvements:
- fix_hyde.py: Replace generic template hyde entries with query-specific ones
using GPT-4o-mini (removed 'comprehensive guide covers everything' pattern)
- fix_lex_filler.py: Remove filler words (overview, tutorial, guide, examples,
documentation, best practices) that were padding rather than genuine search intent
- qmd_expansion_v3.jsonl: Improved dataset with 1,498 high-quality entries
Training data preparation:
- convert_to_chatml.py: Convert to ChatML format for LFM2.5 training
- verify_data.py: Validation script to ensure data quality
- train-lfm2/: Ready-to-use training data (90/10 train/val split)
Data quality metrics:
- 100% success rate (all entries properly formatted)
- Query length: 6-65 chars (avg: 29.3)
- Response length: 307-777 chars (avg: 539.5)
- All entries contain lex, vec, and hyde expansions
- 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
* 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>
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>
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>
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>
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>
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 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>
