qwen35-counterfactual-sft-v12

Qwen3.5-9B Counterfactual Bridge SFT v12

The original Joint Boundary+Causal v12 SFT checkpoint from the 2026-08-21 counterfactual/work-order SFT experiment, arm joint-boundary-causal-v12-all-base-lr2e-8. This is SFT itself, not the later v12-to-SOLO-RL sweep or SOLO RL ckpt174.

Historical training identity

• Initialized independently from Qwen/Qwen3.5-9B, revision c202236235762e1c871ad0ccb60c8ee5ba337b9a; not continued from SFT v5.
• 172 boundary actions plus 24 Gemini-selected causal actions, 196 rows total.
• The causal actions come from 14 exact-state MiniMax-rescue/Qwen-failure pairs. Twenty-three of those actions follow a teacher-generated prefix, rather than occurring exactly at the takeover boundary.
• Joint full-parameter language-model SFT, one optimizer update, constant LR 2e-8, sample-normalized loss, thinking-off data.
• Training-data SHA256: 935eff2516d2fce8618063bdc6865d584c682fdb5aa5d8471e033ddb5dab1111.

Recovered experiment code and report are a historical source snapshot, not a verified restoration of the original Git commit. The training data is not included in this model repository.

Exact archive, not a new conversion

The four original safetensors shards, index, config and tokenizer are published byte-for-byte from the verified durable backup. The index contains 427 trained language-model tensors. There are no vision tensors in this historical export; no frozen base/vision tensors were added or synthesized for upload. The historical model config is preserved. This repository is not a validated complete multimodal checkpoint; a loader must handle the exported text-only weight scope appropriately.

artifact-provenance.json and SHA256SUMS identify the archive and its checksums. UPSTREAM_MODEL_CARD.md preserves the base model card; its benchmark claims do not describe this SFT checkpoint. Optimizer/RNG/training-resume state is absent.

Download

Initially private; use a currently authorized Hugging Face account/token. Pin the commit revision supplied with the publication result.

hf download CharlieLLL/qwen35-counterfactual-sft-v12 --local-dir ./sft-v12
cd sft-v12
sha256sum -c SHA256SUMS

No new GPU-loading or quality evaluation was performed for publication. Formal SWE-bench comparisons require native x86_64 Docker, matched tasks, harness/evaluator hashes and model settings, and valid canonical verifier outcomes with infrastructure errors accounted for. ARM64 diagnostics are not model-quality evidence.