Hugging Face's logo

SaylorTwift
/
lfm-opencode-lora-rl-code

Model card Files
xet
 Community
lfm-opencode-lora-rl-code

Commit History

2 days ago

  1. TASK.md: handoff for the next agent (state, artifacts, outstanding work, sharp edges)

    SaylorTwift HF Staff
    Verified signature
    7a5960d

  2. run_eval.sh: test split only + Jobs-proxy path (missing train dir killed the env server)

    SaylorTwift HF Staff
    Verified signature
    bc532e5

3 days ago

  1. Eval cells now record turns/tool_calls/completion+prompt tokens; summary gets efficiency-vs-base ratios

    SaylorTwift HF Staff
    Verified signature
    371ff91

  2. Resilient worker (transient CaptureContractError -> unscorable, systematic still fatal), resume-from-repo, eval worker base-cells download fix

    SaylorTwift HF Staff
    Verified signature
    e7168a9

4 days ago

  1. Eval robustness: fresh HarborEnv client per rollout (openenv reconnect bug); failed checkpoint evals retried next poll

    SaylorTwift HF Staff
    Verified signature
    75a9ca0

  2. RECIPE: document per-checkpoint eval worker

    SaylorTwift HF Staff
    Verified signature
    b6f8fe5

  3. Add per-checkpoint eval worker (eval_watch.py + run_eval_watch.sh); ONLY_MODEL mode in eval_adapter.py

    SaylorTwift HF Staff
    Verified signature
    0298809

  4. Add RECIPE.md: shareable setup + run recipe with pitfalls

    SaylorTwift HF Staff
    Verified signature
    af7d04e

  5. Fix seam patch: target _opencode block by unique indent (3 apiKey seams exist)

    SaylorTwift HF Staff
    Verified signature
    231d3f3

  6. Patch opencode seam: session id as x-api-key too (Jobs proxy strips Authorization)

    SaylorTwift HF Staff
    Verified signature
    e5f301b

7 days ago

  1. v2: tunnel-free proxy via SPACE_HOST+Jobs expose_public, heartbeat watchdog (30min stall -> fail), keep UploadCheckpointsCallback

    SaylorTwift HF Staff
    Verified signature
    81a1081

8 days ago

  1. Add UploadCheckpointsCallback (transformers v5 removed PushToHubCallback; periodic saves were local-only)

    SaylorTwift HF Staff
    Verified signature
    4e22c33

  2. eval: pass purpose=eval (server requires it with eval_sampling)

    SaylorTwift HF Staff
    Verified signature
    e7a4c0e

  3. Add held-out eval (base vs adapter) + standalone eval job + final eval stage in training job

    SaylorTwift HF Staff
    Verified signature
    c511323

  4. vLLM: --enable-auto-tool-choice --tool-call-parser lfm2 (OpenCode sends tool_choice=auto)

    SaylorTwift HF Staff
    Verified signature
    f4740f8

  5. Set LFM2.5 response_template directly (chat template not recognized by add_response_schema)

    SaylorTwift HF Staff
    Verified signature
    1c2fa85

  6. Pass train_dataset (required by AsyncGRPOTrainer ctor in 1.15.0)

    SaylorTwift HF Staff
    Verified signature
    2a4d2a5

  7. Add kernels==0.16.0 (trl 1.15.0 loads kernels-community/flash-attn3)

    SaylorTwift HF Staff
    Verified signature
    8067c88

  8. Pin trl==1.15.0 (pre-tree-packing, LFM2 hybrid compatible) + vllm v0.31.0 image

    SaylorTwift HF Staff
    Verified signature
    73712c7

  9. Install git in job (vllm image lacks it)

    SaylorTwift HF Staff
    Verified signature
    9cd9f57

  10. Use python3 (vllm image has no python alias)

    SaylorTwift HF Staff
    Verified signature
    ea9b383

  11. Switch sandboxes to Harbor hf-sandbox backend; add pinned task prep with docker_image patch; dataset discovery via prompt_rows()

    SaylorTwift HF Staff
    Verified signature
    a7525f2

  12. OpenCode-only LoRA RL on LFM2.5-2.6B (verified against trl 1.16.0 / openenv 0.10.0)

    SaylorTwift HF Staff
    Verified signature
    f0273c6

  13. initial commit

    SaylorTwift HF Staff
    Verified signature
    2b008b7