fine-tuning-4b-code-model-training-auxiliary-targets-name-each CLM-0568mechanismsingle paperReviewed by AI
When fine-tuning a 4B code model, training on auxiliary targets that name each program region's correctness obligations (derived from aligning an AST implementation graph with Lean proof-flow traces) raises solved counts over code-only SFT, but the auxiliary-target stage alone shifts the model off the Python-only output distribution and must be followed by a code-only consolidation stage to avoid losing broad function-level ability.
Capability: Generating and editing working code
Observed on
Qwen3.5-4B-Base with LoRA adapters, greedy Pass@1, LiveCodeBench v6 and BigCodeBench; benchmark reference solutions used as training targets..
Sources
- Validated by: Pass@1 solved counts against base model, code-only SFT and CodeRL on LiveCodeBench v6 (50 vs 38 CodeRL), BigCodeBench Hard (23 vs 16) and BigCodeBench Full (363 vs 359); ablation shows injection alone reaches 44 on LCB v6 but drops BCB Full from 354 to 339, and consolidation-only reaches 49/352/19.. Measured on three benchmarks with one 4B backbone, single greedy sample, no variance reported and no matched-compute control for the extra training steps; [truncated]
Either way, bring evidence: a link, or what you saw.
Disagreeing is the most useful thing you can do here. Nothing on this page waits for a reviewer: the counts above move when evidence lands, and every source stays attached with who brought it.
Related claims
- For 4B-8B open base models mid-trained for tool use, grounding-oriented supervision (documentation/code turned into QA) and execution-oriented supervision (executable trajectories from real APIs and MCP skills) help different things, and only the combination improves everything: trajectories alone give the bigger function-calling gain while grounding alone transfers better to interactive and unseen-tool benchmarks, and each branch alone scores below no mid-training on MCP-Universe.Using the tools it is given · reviewed by AI
- When injecting new or counterfactual facts into a 3-4B instruct model by post-training, supervised fine-tuning on extracted QA pairs binds facts to their training question form — recall of the single fact stays reasonable while multi-source and inferential questions over the same facts collapse — and adding more SFT epochs does not close the gap, whereas mixed-policy RL that supplies the golden answer as an off-policy trajectory when on-policy rollouts fail recovers much of the generalization.Absorbing and flexibly using new knowledge · reviewed by AI
- When fine-tuning 7-8B code models on an LLM-synthesized Verilog instruction corpus, training on a 20-25% subset selected for functional correctness, instruction-solution alignment, and coverage of Verilog structural clusters (AST, CFG, netlist features) yields higher Pass@1 than training on the full corpus, because the automated corpus carries noise and redundancy that filtering removes.Generating and editing working code · reviewed by AI
- For a 7B code model fine-tuned on competitive-programming tasks, training it to emit a fixed-schema requirement specification (goals, input/output constraints, examples, edge cases, dependencies) before the code raises Pass@1 more than training it to emit free-form reasoning or a stepwise implementation plan of similar length, with the largest margin on an out-of-distribution benchmark.Generating and editing working code · reviewed by AI
- For a domain-specific language essentially absent from pretraining corpora (FORM symbolic algebra), QLoRA fine-tuning an 8B model on a few thousand examples validated by executing them through the language's own interpreter beats frontier open-weight models of up to 756B parameters given syntax documentation in context, on execution rate and strict output match for the larger benchmarks, while general reasoning scores drop under 3 points.Generating and editing working code · reviewed by AI
Notes
Drafted from the paper by a model and filed unreviewed. Visible here so it can be read, not because anyone has vouched for it: it does not move any technique's standing and does not count toward the internal scorecard. Drafted confidence: low. Falsifier as drafted: Running the injection stage alone on other backbones shows no drop in broad function-level pass rate, or a matched-compute code-only SFT run with the same total steps matches the full two-stage pipeline's solved counts.